The paper presents a new combinatorial algorithm for the Correlation Clustering problem that achieves a significantly better approximation factor than the classic 3-approximation.
Key highlights:
The algorithm uses a local search approach, where at each iteration it tries to swap in a cluster (i.e., a set of vertices) to improve the current clustering.
To escape bad local minima, the algorithm introduces a "flipping" technique - it doubles the weight of the edges cut by the current solution, and then runs the local search again on the modified instance.
The authors show that by iterating this local search and flipping process, the algorithm achieves a 2 - 2/13 < 1.847-approximation, a drastic improvement over the previous 3-approximation.
The authors also provide efficient implementations of the local search algorithm in various computational models, including sublinear time, streaming, and massively parallel computation (MPC), while preserving the improved approximation guarantee.
The key technical ingredients are: (i) a careful analysis of the local search algorithm and the role of the flipping technique, (ii) a preclustering step to enable efficient implementations, and (iii) novel sampling and aggregation techniques to estimate the cost of potential swaps.
Overall, the paper presents a significant advancement in the approximability of the Correlation Clustering problem, both in terms of the approximation factor and the efficiency of the algorithms across different computational models.
To Another Language
from source content
arxiv.org
Key Insights Distilled From
by Vincent Cohe... at arxiv.org 04-09-2024
https://arxiv.org/pdf/2404.05433.pdfDeeper Inquiries